Integrating skeleton based representations for robust yoga pose classification using deep learning models

Clinical Snapshot

Key Findings

• P Value: Not reported

• Effect Size: Best performance: VGG16 with MediaPipe Pose skeleton input achieved 96.09% classification accuracy; skeleton-based inputs consistently outperformed raw image inputs across all architectures

• Primary Outcome: Yoga pose classification accuracy across 16 pose classes using three deep learning architectures and three input modalities

• Nnt Or Sensitivity: Per-class sensitivity and specificity not reported; no NNT applicable (non-clinical computational study); Grad-CAM used for qualitative model interpretability only

• Confidence Interval: Not reported

Clinical Application

MediaPipe Pose is an open-source, computationally lightweight framework suitable for mobile and web deployment. VGG16 is a well-established architecture with broad hardware support. However, real-time deployment feasibility, latency benchmarks, and integration into clinical workflows have not been assessed in this study. Yoga is widely practised in Australia, with growing integration into allied health and physiotherapy-led rehabilitation programs. The TGA does not currently regulate software-based pose classification tools as medical devices unless they make diagnostic or therapeutic claims, though this may change under evolving SaMD (Software as a Medical Device) frameworks. The RACGP does not have specific guidelines for AI-assisted yoga pose correction. PBS listing is not applicable. Any clinical deployment in Australia would require prospective validation on diverse Australian populations, AHPRA-compliant governance, and likely TGA SaMD assessment if injury prevention claims are made. The dataset's origin from Bangladesh limits direct demographic applicability to Australian populations, which are more ethnically diverse and may include practitioners with different body habitus distributions. Theoretically applicable to yoga practitioners seeking automated pose feedback, physiotherapy rehabilitation settings using yoga-based movement, and consumer health applications. However, no clinical population has been prospectively evaluated.

Abstract

Yoga is a popular form of exercise worldwide due to its spiritual and physical health benefits, but incorrect postures can lead to injuries. Automated yoga pose classification has therefore gained importance to reduce reliance on expert practitioners. While human pose keypoint extraction models have shown high potential in action recognition, systematic benchmarking for yoga pose recognition remains limited, as prior works often focus solely on raw images or a single pose extraction model. In this study, we introduce a curated dataset, "Yoga-16", which addresses limitations of existing datasets, and systematically evaluate three deep learning architectures-VGG16, ResNet50, and Xception-using three input modalities: direct images, MediaPipe Pose skeleton images, and YOLOv8 Pose skeleton images. Our experiments demonstrate that skeleton-based representations outperform raw image inputs, with the highest accuracy of 96.09% achieved by VGG16 with MediaPipe Pose skeleton input. Additionally, we provide interpretability analysis using Grad-CAM, offering insights into model decision-making for yoga pose classification with cross validation analysis.

References

1. 1.Mohiuddin, M., Hossain, S. M. M., Khanam, S., Barua, P., Barua, A., & Hossain, M. D. T. (2026). Integrating skeleton based representations for robust yoga pose classification using deep learning models. Scientific Reports. https://doi.org/10.1109/LSENS.2022.3145750 [Note: DOI-journal attribution inconsistency identified; readers should verify source independently via PubMed ID 42449118]